Zero-Shot Classification
Transformers
Safetensors
English
qwen3_5
feature-extraction
legal
decision-model
system-one
classification
citator
case-law
custom_code
Instructions to use jonathankleiman/lev with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use jonathankleiman/lev with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("zero-shot-classification", model="jonathankleiman/lev", trust_remote_code=True)# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModel processor = AutoProcessor.from_pretrained("jonathankleiman/lev", trust_remote_code=True) model = AutoModel.from_pretrained("jonathankleiman/lev", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
Model card, figures and technical report
Browse files- .gitattributes +5 -0
- README.md +213 -0
- assets/banner.png +0 -0
- assets/calibration_gold.png +3 -0
- assets/confusion_treatment_lev.png +3 -0
- assets/headline_gold.png +0 -0
- assets/headline_silver.png +0 -0
- assets/latency.png +0 -0
- assets/per_task_gold.png +3 -0
- assets/per_task_silver.png +3 -0
- lev-technical-report.pdf +3 -0
.gitattributes
CHANGED
|
@@ -34,3 +34,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
assets/calibration_gold.png filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
assets/confusion_treatment_lev.png filter=lfs diff=lfs merge=lfs -text
|
| 39 |
+
assets/per_task_gold.png filter=lfs diff=lfs merge=lfs -text
|
| 40 |
+
assets/per_task_silver.png filter=lfs diff=lfs merge=lfs -text
|
| 41 |
+
lev-technical-report.pdf filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,213 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
library_name: transformers
|
| 6 |
+
pipeline_tag: zero-shot-classification
|
| 7 |
+
base_model: vllm-sr/d3-mini
|
| 8 |
+
tags:
|
| 9 |
+
- legal
|
| 10 |
+
- decision-model
|
| 11 |
+
- system-one
|
| 12 |
+
- classification
|
| 13 |
+
- citator
|
| 14 |
+
- case-law
|
| 15 |
+
- custom_code
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+

|
| 19 |
+
|
| 20 |
+
# Lev
|
| 21 |
+
|
| 22 |
+
Lev is a 4B decision model for case law. You give it an excerpt of a court decision and one or more typed questions, and it returns a probability for every answer option. It does not generate text. It was built for citator work: deciding how a decision treats a case it cites, whether the cited case is an earlier stage of the same proceeding, and what happened to it on appeal. It also answers ten other classification questions about decisions and passages.
|
| 23 |
+
|
| 24 |
+
Lev is a fine-tune of [d3-mini](https://huggingface.co/vllm-sr/d3-mini) (Qwen3.5-4B backbone, Apache-2.0) and loads with the same runtime.
|
| 25 |
+
|
| 26 |
+
| | |
|
| 27 |
+
| --- | --- |
|
| 28 |
+
| **Parameters** | 4.54B (the d3-mini checkpoint with merged adapters and a tuned readout) |
|
| 29 |
+
| **Input** | Text: an excerpt of a decision, up to a few thousand tokens |
|
| 30 |
+
| **Questions** | Choice (named options with definitions) and yes/no |
|
| 31 |
+
| **Trained for** | 13 questions on Canadian court decisions in English |
|
| 32 |
+
| **License** | Apache-2.0 |
|
| 33 |
+
|
| 34 |
+
## Results
|
| 35 |
+
|
| 36 |
+
Scores are on a held-out test set of 1428 excerpts from decisions dated after the training data ends, with 5367 judgments in all. Every system got the same excerpt, instructions and option definitions. The gold set is the 4674 judgments on which two independent annotators agreed; the silver set is every judgment, labelled by the first annotator only. Both annotators are models, not people. Details and caveats are in the [technical report](lev-technical-report.pdf).
|
| 37 |
+
|
| 38 |
+
**Gold (two annotators agree), accuracy %**
|
| 39 |
+
|
| 40 |
+
| Task | n | **Lev** | Jev 1.13 | OpenAI Decisions (gpt-6-luna) | d3-mini (base) |
|
| 41 |
+
|---|---|---|---|---|---|
|
| 42 |
+
| Treatment of the cited case | 407 | **82.8** | 79.9 | 69.0 | 68.8 |
|
| 43 |
+
| Appeal history | 590 | 93.2 | **93.9** | 86.4 | 86.4 |
|
| 44 |
+
| Reported negative history | 596 | 95.6 | **97.7** | 95.8 | 91.9 |
|
| 45 |
+
| Cited in dissent | 598 | 96.0 | 95.5 | **97.0** | 83.8 |
|
| 46 |
+
| *Mean of the four citator tasks* | 2191 | **91.9** | 91.7 | 87.1 | 82.7 |
|
| 47 |
+
| *Mean of seven decision-level tasks* | 2123 | 94.8 | 95.8 | **96.0** | 91.5 |
|
| 48 |
+
| *Mean of two passage-level tasks* | 360 | **89.3** | 87.9 | 85.6 | 83.1 |
|
| 49 |
+
| *Mean of all 13* | 4674 | 93.1 | **93.4** | 91.7 | 87.5 |
|
| 50 |
+
|
| 51 |
+
**Silver (first annotator), accuracy %**
|
| 52 |
+
|
| 53 |
+
| Task | n | **Lev** | Jev 1.13 | OpenAI Decisions (gpt-6-luna) | d3-mini (base) |
|
| 54 |
+
|---|---|---|---|---|---|
|
| 55 |
+
| Treatment of the cited case | 623 | **68.7** | 57.8 | 53.1 | 52.2 |
|
| 56 |
+
| Appeal history | 623 | **90.5** | 89.7 | 82.2 | 82.5 |
|
| 57 |
+
| Reported negative history | 623 | 93.1 | **94.2** | 93.4 | 88.4 |
|
| 58 |
+
| Cited in dissent | 623 | **94.4** | 92.6 | 94.1 | 82.5 |
|
| 59 |
+
| *Mean of the four citator tasks* | 2492 | **86.7** | 83.6 | 80.7 | 76.4 |
|
| 60 |
+
| *Mean of seven decision-level tasks* | 2415 | **89.7** | 89.0 | 89.2 | 84.9 |
|
| 61 |
+
| *Mean of two passage-level tasks* | 460 | **76.1** | 70.9 | 67.9 | 70.4 |
|
| 62 |
+
| *Mean of all 13* | 5367 | **86.7** | 84.5 | 83.3 | 80.1 |
|
| 63 |
+
|
| 64 |
+

|
| 65 |
+
|
| 66 |
+
On the gold set Lev and Jev are level over all 13 tasks (93.1% and 93.4%; the 95% interval on the difference is -1.2 to +0.3 points), and Lev is ahead of the OpenAI Decisions API by 1.6 points (+0.7 to +2.6). On the silver set Lev is ahead of both, by 2.2 and 3.7 points. Silver labels come from the same annotator as Lev's training labels, so that margin flatters it; the gold comparison is the fairer one.
|
| 67 |
+
|
| 68 |
+
Where Lev is better and worse:
|
| 69 |
+
|
| 70 |
+
- **Citation treatment** is its best task: 82.8% on gold against 79.9% (Jev) and 69.0% (OpenAI), with macro-F1 of 70.5% against 60.9% and 60.2%.
|
| 71 |
+
- It assigns a treatment to a case that is only quoted or only argued by a party less often than the hosted systems (7.7% of such citations on gold, against 13.8% and 12.3%; there are only 65 of them, so treat the ordering as tentative).
|
| 72 |
+
- It is behind on **area of law** (85.7% against 93.2% and 92.5%). For that question alone, use a hosted model.
|
| 73 |
+
- **Negative treatment needs a threshold.** If you take the top answer as the flag, Lev catches 71% of the 38 negative treatments in the gold set, against 97% for Jev and 84% for OpenAI. If you flag a citation whenever the probabilities of `questioned`, `not_followed`, `overruled` and `superseded_by_statute` sum to 0.2 or more, it catches 95% at 88% precision (Jev 100% at 68%, OpenAI 97% at 90%). Do not use Lev's top answer as a negative-treatment flag.
|
| 74 |
+
|
| 75 |
+
```python
|
| 76 |
+
NEGATIVE = ("questioned", "not_followed", "overruled", "superseded_by_statute")
|
| 77 |
+
p = result["answers"]["treatment"]["probabilities"]
|
| 78 |
+
flag_for_review = sum(p[k] for k in NEGATIVE) >= 0.2
|
| 79 |
+
```
|
| 80 |
+
|
| 81 |
+
### Public, human-labelled tasks
|
| 82 |
+
|
| 83 |
+
None of these tasks were used in training. The Australian set has treatment labels written by the deciding judges; the others are LegalBench tasks about judicial decisions.
|
| 84 |
+
|
| 85 |
+
| Task | n | **Lev** | Jev 1.13 | OpenAI Decisions (gpt-6-luna) | d3-mini (base) |
|
| 86 |
+
|---|---|---|---|---|---|
|
| 87 |
+
| Treatment, Federal Court of Australia (judge-labelled) | 818 | **42.8** | 40.8 | 38.0 | 40.0 |
|
| 88 |
+
| Overruling (LegalBench) | 300 | 95.0 | 93.3 | 92.0 | **95.3** |
|
| 89 |
+
| Tax Court of Canada outcomes (LegalBench) | 244 | 98.0 | **99.2** | 96.7 | 98.4 |
|
| 90 |
+
| Function of decision section (LegalBench) | 300 | **47.3** | 44.0 | 45.7 | 46.7 |
|
| 91 |
+
| Oral argument question purpose (LegalBench) | 300 | 45.0 | 44.3 | **47.7** | 39.7 |
|
| 92 |
+
| SCALR holdings (LegalBench) | 300 | 76.3 | 79.7 | **83.6** | 75.0 |
|
| 93 |
+
| *Mean of six* | | **67.4** | 66.9 | 67.3 | 65.8 |
|
| 94 |
+
|
| 95 |
+
Fine-tuning did not cost Lev its general ability: its mean over the six tasks is 67.4%, against 65.8% for the base model, and all four systems are within two points of each other. The citator skill transfers to Australia only slightly (42.8% against 40.0% for the base model, a difference that is not statistically reliable). Every system is right about 86% of the time on whether the Australian court limited or rejected the case, and near chance on which positive label the judge chose. The report discusses why.
|
| 96 |
+
|
| 97 |
+
## Quickstart
|
| 98 |
+
|
| 99 |
+
```bash
|
| 100 |
+
pip install "transformers==5.17.0" torch safetensors accelerate
|
| 101 |
+
pip install flash-linear-attention # optional, faster linear-attention layers on GPU
|
| 102 |
+
```
|
| 103 |
+
|
| 104 |
+
```python
|
| 105 |
+
from transformers import AutoModel
|
| 106 |
+
|
| 107 |
+
model = AutoModel.from_pretrained("jfloydian/lev", trust_remote_code=True)
|
| 108 |
+
|
| 109 |
+
excerpt = """Citing decision: Smith v. Jones, 2025 ONCA 100 | Court of Appeal for Ontario (appeal court, Ontario) | decided 2025-02-03
|
| 110 |
+
Cited case: Brown v. Green, 2010 ONSC 55
|
| 111 |
+
|
| 112 |
+
Paragraphs that cite the case, with neighbours:
|
| 113 |
+
[41] The respondent relies on Brown v. Green. With respect, that decision cannot stand with the later
|
| 114 |
+
guidance of this court in Grey, and I decline to follow it.
|
| 115 |
+
[42] The governing test is the one stated in Grey at para. 17."""
|
| 116 |
+
|
| 117 |
+
result = model.system_one(
|
| 118 |
+
state=excerpt,
|
| 119 |
+
questions={
|
| 120 |
+
"treatment": {
|
| 121 |
+
"type": "choice",
|
| 122 |
+
"instructions": "How does the citing decision treat the cited case? If it treats the case in more than one way, choose the most negative treatment.",
|
| 123 |
+
"criteria": {
|
| 124 |
+
"followed": "The court treats the cited case as binding or adopts its rule as the governing rule, and decides the point with it.",
|
| 125 |
+
"applied": "The court applies the cited case's rule or test to the facts before it.",
|
| 126 |
+
"distinguished": "The court says the cited case does not govern because the facts or the law differ.",
|
| 127 |
+
"not_followed": "The court declines to follow the cited case.",
|
| 128 |
+
"referred_to": "The court cites the case in its own reasoning for a point, without more.",
|
| 129 |
+
},
|
| 130 |
+
},
|
| 131 |
+
"dissent": {"type": "noul", "instructions": "Do dissenting reasons in the citing decision cite or treat the cited case?"},
|
| 132 |
+
},
|
| 133 |
+
)
|
| 134 |
+
print(result["answers"]["treatment"]["choice"], result["answers"]["treatment"]["probabilities"])
|
| 135 |
+
```
|
| 136 |
+
|
| 137 |
+
The 13 questions Lev was trained on, with the exact instructions and option definitions, are in [`questions.py`](questions.py). Use that wording for the trained tasks. You can also ask your own questions with your own options, as with d3-mini; the public-task table above gives a sense of how far that carries.
|
| 138 |
+
|
| 139 |
+
### How to build the excerpt
|
| 140 |
+
|
| 141 |
+
The model was trained on excerpts with a fixed layout, and it does best when you keep to it.
|
| 142 |
+
|
| 143 |
+
- For a citation pair: a line naming the citing decision, its court and date; a line naming the cited case; up to 500 characters of front matter; the first paragraph; the paragraphs that cite the case with one neighbour on each side; the last two paragraphs.
|
| 144 |
+
- For a decision-level question: the header line, up to 900 characters of front matter, the first three paragraphs and the last five.
|
| 145 |
+
- For a passage-level question: the header line, the issue or the test element, and the paragraphs that deal with it.
|
| 146 |
+
|
| 147 |
+
[`excerpts.py`](excerpts.py) has the functions used to build all three.
|
| 148 |
+
|
| 149 |
+
## The questions
|
| 150 |
+
|
| 151 |
+
| Group | Question | Options |
|
| 152 |
+
| --- | --- | --- |
|
| 153 |
+
| Citator | `treatment` | followed, applied, considered, distinguished, questioned, not_followed, overruled, superseded_by_statute, referred_to, inside_quotation, other_court_or_party, same_matter, not_named |
|
| 154 |
+
| Citator | `history` | not_same_matter, affirmed, reversed, reversed_in_part, varied, set_aside, restored, remitted, earlier_ruling |
|
| 155 |
+
| Citator | `reported` | none, overruled, reversed, superseded_by_statute, not_followed, doubted |
|
| 156 |
+
| Citator | `dissent` | yes / no |
|
| 157 |
+
| Decision | `outcome` | 16 dispositions |
|
| 158 |
+
| Decision | `decision_type` | 10 kinds of document |
|
| 159 |
+
| Decision | `proceeding` | 11 kinds of proceeding |
|
| 160 |
+
| Decision | `area` | 49 areas of law |
|
| 161 |
+
| Decision | `publication_ban`, `self_represented`, `separate_opinions` | yes / no |
|
| 162 |
+
| Passage | `standard_of_review` | correctness, reasonableness, palpable_and_overriding_error, patent_unreasonableness, other, not_stated |
|
| 163 |
+
| Passage | `element_finding` | met, not_met, not_decided |
|
| 164 |
+
|
| 165 |
+
Three of the treatment options exist to keep a citator honest. `inside_quotation` is for a case that appears only in text the court is quoting. `other_court_or_party` is for a case that appears only in a party's argument or another court's reasoning. `reported` is a separate question for what the court says someone else did to the case. A model that reads all of these as the court's own treatment reports things that did not happen.
|
| 166 |
+
|
| 167 |
+
## Calibration
|
| 168 |
+
|
| 169 |
+
A single temperature (1.2) was fitted on held-out data drawn at natural label frequencies and is stored in `decision_config.json`. Averaged over the 13 tasks, expected calibration error on the gold set is 2.9% for Lev, 4.1% for Jev and 3.7% for OpenAI. Answers Lev gives with at least 90% confidence are right 99% of the time; in the 80 to 90% band, 89%. Ranking its citator answers by confidence, Lev can decide 89% of the gold judgments before accuracy on the accepted set drops below 98% (Jev 87%, OpenAI 73%).
|
| 170 |
+
|
| 171 |
+

|
| 172 |
+
|
| 173 |
+
## Training
|
| 174 |
+
|
| 175 |
+
Lev was trained on a private corpus of Canadian court decisions with model-assisted labels; test labels were adjudicated by a second, independent model. Training decisions are dated before 1 July 2024 and test decisions on or after it.
|
| 176 |
+
|
| 177 |
+
- Low-rank adapters (rank 32) on every linear layer of the upper 16 of 32 decoder layers, plus the readout, trained through a reimplementation of the d3 forward pass, since the released runtime is inference-only.
|
| 178 |
+
- Three stages, 89,718 training instances in all: a class-balanced round split over three GPUs and averaged, a second class-balanced round, and a short final stage on data sampled with square-root label frequencies so that the model stops over-predicting rare answers.
|
| 179 |
+
- The order of the options is shuffled for every training example, so the model has to read the definitions.
|
| 180 |
+
- Adapters are merged into the bfloat16 weights. The merged model scored the same as the unmerged one on held-out data.
|
| 181 |
+
|
| 182 |
+
The adapters and the code are in [jfloydian/lev-lora](https://huggingface.co/jfloydian/lev-lora), and the [technical report](lev-technical-report.pdf) has the details.
|
| 183 |
+
|
| 184 |
+
## Limitations
|
| 185 |
+
|
| 186 |
+
- **Both annotators of the main test set are models.** "Gold" means two independent model annotators agree. No person reviewed the labels. The public tasks above have human labels.
|
| 187 |
+
- **The test set over-samples rare labels.** The scores describe how each system handles each kind of case, not how often it will be right on a random citation.
|
| 188 |
+
- **Canadian decisions in English with numbered paragraphs.** Other jurisdictions, French-language decisions and differently built excerpts are outside what was tested, apart from the public tasks above.
|
| 189 |
+
- **Excerpts hide things.** Lev sees the paragraphs around a citation, not the decision. Positive treatment in particular is often only visible from how the rest of the decision uses the case, and two careful readers of the same decision agree that a case was followed or applied only about half the time.
|
| 190 |
+
- **It classifies an excerpt; it is not a citator.** It does not find citations, resolve them to cases, or check that the excerpt is complete. Treat its output as a lead to check against the decision. It is not legal advice.
|
| 191 |
+
- **Images and video** are supported by the d3 runtime it inherits, but Lev was trained and tested on text only.
|
| 192 |
+
- **Hosted baselines change.** Jev 1.13 and the OpenAI Decisions API (in public beta) were called on 12 October 2026.
|
| 193 |
+
|
| 194 |
+
## Files
|
| 195 |
+
|
| 196 |
+
- `model-0000{1,2}-of-00002.safetensors`, `readout.safetensors`, `decision_config.json`: the model in d3 package format.
|
| 197 |
+
- `d3_*.py`, `modeling_d3.py`, `pipeline_d3.py`: the d3 runtime, unchanged from d3-mini.
|
| 198 |
+
- `questions.py`, `excerpts.py`: the trained questions and the excerpt builders.
|
| 199 |
+
- `lev-technical-report.pdf`: the technical report.
|
| 200 |
+
- The LoRA adapters and the training and evaluation code are in [jfloydian/lev-lora](https://huggingface.co/jfloydian/lev-lora).
|
| 201 |
+
|
| 202 |
+
## License and attribution
|
| 203 |
+
|
| 204 |
+
Apache-2.0. Lev is built on d3-mini by the vLLM Semantic Router team, which is built on Qwen3.5-4B; their licence and notice files are included. If you use Lev, please credit Jonathan Kleiman.
|
| 205 |
+
|
| 206 |
+
```bibtex
|
| 207 |
+
@misc{kleiman2026lev,
|
| 208 |
+
title = {Lev: a 4B open decision model for citator and case-law classification},
|
| 209 |
+
author = {Kleiman, Jonathan},
|
| 210 |
+
year = {2026},
|
| 211 |
+
howpublished = {\url{https://huggingface.co/jfloydian/lev}}
|
| 212 |
+
}
|
| 213 |
+
```
|
assets/banner.png
ADDED
|
assets/calibration_gold.png
ADDED
|
Git LFS Details
|
assets/confusion_treatment_lev.png
ADDED
|
Git LFS Details
|
assets/headline_gold.png
ADDED
|
assets/headline_silver.png
ADDED
|
assets/latency.png
ADDED
|
assets/per_task_gold.png
ADDED
|
Git LFS Details
|
assets/per_task_silver.png
ADDED
|
Git LFS Details
|
lev-technical-report.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cb886bc5ec0e6a2dedef5e35308e6a06d44362d164b1a97152c9b86289065376
|
| 3 |
+
size 389472
|